A solo RTX 5090 image-model run broke in absurd ways before it started working
Creative-Listen-6847 · reddit · 2026-07-23
A Reddit creator documents training an image model from scratch on a single RTX 5090, starting with about 37k image-caption pairs and iterating through multiple broken runs.
What went wrong
- The first model collapsed into a “snow-only” generator because it only learned a blurry caption summary.
- A copy-paste bug caused the trainer to overwrite the live model with an older backup every time it saved previews, erasing roughly 92,000 steps.
- The larger run hit practical failures instead of model failures: the data loader crashed when opening 71 files at once, dataset building ran out of memory and left 47 GB of junk files, a single checkmark character crashed training, and one missing backslash broke the launch script.
What eventually worked
- After rebuilding the architecture to read captions word by word and scaling the dataset, the model started producing increasingly correct images.
- Examples included “red dress” yielding a red dress, “white cat” producing a white blob with the right shape, and “red sports car” evolving from a jar-like object into an actual car.
- The creator notes that the loss curve barely moved while image quality clearly improved, showing that loss alone was a poor signal for this setup.
More from Multimodal
- A Midjourney prompt recreates 1970s New York street photography — tisch_eins · 2026-07-23
- Microsoft opens Mage-Flow, a 4B native-resolution model for image generation and editing — pmttyji · 2026-07-23
- Flint Image offers four image outputs per prompt at 67% lower cost — miilesus · 2026-07-23
- 3 Prompts to an Infinite Universe: Showcasing an AI Game Dev Workflow — scaling01 · 2026-07-23
- Open-Source Community May Welcome a SeeDance-Level Model — cocktailpeanut · 2026-07-23
- A Midjourney prompt formula for dramatic graphic novel covers — tisch_eins · 2026-07-23